Online Imbalanced Learning with Kernels

Imbalanced learning, or learning from imbalanced data, is a challenging problem in both academy and industry. Nowadays, the streaming imbalanced data become popular and trigger the volume, velocity, and variety issues of learning from these data. To tackle these issues, online learning algorithms are proposed to learn a linear classifier via maximizing the AUC score. However, the developed linear classifiers ignore the learning power of kernels. In this paper, we therefore propose online imbalanced learning with kernels (OILK) to exploit the non-linearity and heterogeneity embedded in the imbalanced data. Different from previously proposed work, we optimize the AUC score to learn a non-linear representation via the kernel trick. To relieve the computational and storing cost, we also investigate different buffer update policies, including first-in-first-out (FIFO) and reservoir sampling (RS), to maintain a fixed budgeted buffer on the number of support vectors. We demonstrate the properties of our proposed OILK through detailed experiments.